[foxhill]

i am quite familiar with compiler internals :)

of course, across the TU boundary things get difficult, but i don’t think it’s fair to dismiss LTO entirely (although.. i agree with the thesis that it’s not particularly.. good)

similarly, de-virtualisation is an optimisation technique compilers will aggressively use to improve performance, although you’re right that it can’t look through another source file, so it is not without limitations here.

but we’re not being general, we’re being specific; the safety issues that are being discussed are well within the remit of the c++ type system here, and i don’t think we’re doing any favours to anyone by letting this rvalue be accessed in this way. it is certainly not idiomatic to provide library code that can so violently implode with seemingly regular use. i find it difficult to believe that lifetime issues like this are undiagnosable.